fix gpu decorator outside gradio issue

app.py

@@ -10,7 +10,6 @@ import os
 hf_token = os.getenv("HF_TOKEN")
 
 
-@spaces.GPU
 def launch_app():
 
     @spaces.GPU
@@ -81,42 +80,42 @@ def launch_app():
         def forward(self, base, source=None, subspaces=None):
             return torch.relu(self.proj(base))
 
-    # Load tokenizer and model
-    tokenizer = AutoTokenizer.from_pretrained(model_name, token=hf_token)
-    model = AutoModelForCausalLM.from_pretrained(model_name, device_map='auto', token=hf_token).to("cuda" if torch.cuda.is_available() else "cpu")
-
-    # Load fast model inference pipeline
-    pipe = pipeline(
-        task="text-generation",
-        model=model_name,
-        use_fast=True, 
-        token=hf_token
-    )
-
-    path_to_params = hf_hub_download(
-        repo_id=interpreter_name,
-        filename=interpreter_path,
-        force_download=False,
-    )
-    params = torch.load(path_to_params, map_location="cuda" if torch.cuda.is_available() else "cpu")
-    encoder = Encoder(embed_dim=params.shape[0], latent_dim=params.shape[1]).to("cuda" if torch.cuda.is_available() else "cpu")
-    encoder.proj.weight.data = params.float()
-    pv_model = pv.IntervenableModel({
-        "component": interpreter_component,
-        "intervention": encoder}, model=model).to("cuda" if torch.cuda.is_available() else "cpu")
-
-    # Load dictionary
-    all_concepts = get_concepts_dictionary(dictionary_url)
-
-    description_text = """
-    ## Does an LLM Think Like You?
-    Input a prompt and a concept that you think is most relevant for your prompt. See how much (if at all) the LLM uses that concept when processing your prompt.
-    Examples:
-    - **Prompt**: What is 2+2? **Concept**: math
-    - **Prompt**: I really like anchovies on pizza but I know a lot of people don't. **Concept**: food
-    """
-
     with gr.Blocks() as demo:
+        # Load tokenizer and model
+        tokenizer = AutoTokenizer.from_pretrained(model_name, token=hf_token)
+        model = AutoModelForCausalLM.from_pretrained(model_name, device_map='auto', token=hf_token).to("cuda" if torch.cuda.is_available() else "cpu")
+
+        # Load fast model inference pipeline
+        pipe = pipeline(
+            task="text-generation",
+            model=model_name,
+            use_fast=True, 
+            token=hf_token
+        )
+
+        path_to_params = hf_hub_download(
+            repo_id=interpreter_name,
+            filename=interpreter_path,
+            force_download=False,
+        )
+        params = torch.load(path_to_params, map_location="cuda" if torch.cuda.is_available() else "cpu")
+        encoder = Encoder(embed_dim=params.shape[0], latent_dim=params.shape[1]).to("cuda" if torch.cuda.is_available() else "cpu")
+        encoder.proj.weight.data = params.float()
+        pv_model = pv.IntervenableModel({
+            "component": interpreter_component,
+            "intervention": encoder}, model=model).to("cuda" if torch.cuda.is_available() else "cpu")
+
+        # Load dictionary
+        all_concepts = get_concepts_dictionary(dictionary_url)
+
+        description_text = """
+        ## Does an LLM Think Like You?
+        Input a prompt and a concept that you think is most relevant for your prompt. See how much (if at all) the LLM uses that concept when processing your prompt.
+        Examples:
+        - **Prompt**: What is 2+2? **Concept**: math
+        - **Prompt**: I really like anchovies on pizza but I know a lot of people don't. **Concept**: food
+        """
+
         gr.Markdown(description_text)
         with gr.Row():
             prompt_input = gr.Textbox(label="Enter a prompt", value="I really like anchovies on pizza but I know a lot of people don't.")